RISE gives driving world models an adaptive imagination budget
World Action Models plan by imagining how a driving scene might evolve before choosing an action, but existing methods run this imagination for a fixed length on every scene regardless of how much it is actually needed. A new paper proposes RISE, short for Refining Imagination through SElective Rollout, a system-level framework that instead makes a sequential Roll or Stop decision at each step, based on the expected planning benefit of continuing the rollout. The mechanism has two parts: a Latent Evaluator that estimates the risk revealed by the imagined sequence so far and how much further planning could still improve, and a Rollout Gate that weighs that expected benefit against the extra computation it would cost to keep going.
Because real driving logs only ever record one outcome that actually happened, the authors also built CounterDrive, a counterfactual dataset that adds diverse alternative outcomes and risk levels to enrich the range of future dynamics available for training and to supply localized risk supervision. Every sample kept in CounterDrive goes through expert verification, with annotation of trajectory validity, when an incident begins, and its causal category, which the authors present as a reusable resource for further safety-critical world-modeling research beyond this paper.
The authors report testing RISE on the NAVSIM and nuScenes planning benchmarks, where it achieves the best overall planning performance among the methods compared while cutting unnecessary rollout, and they report additional transfer results showing it works as a plug-in across different World Action Model architectures. The available text does not include the specific accuracy or scoring figures behind these claims, name the paper's authors or institutions, state how large CounterDrive is, or give a timeframe or publication venue for the work.
Key facts
- RISE replaces the fixed imagination budget used by prior World Action Models with a per-step Roll or Stop decision.
- A Latent Evaluator estimates current risk and potential planning improvement, while a Rollout Gate weighs that benefit against added compute cost.
- The authors built CounterDrive, a counterfactual driving dataset with diverse outcomes and risk levels, because real driving logs only capture one realized future.
- Each CounterDrive sample undergoes expert verification with annotation of trajectory validity, incident onset, and causal category.
- On the NAVSIM and nuScenes benchmarks, RISE achieves the best overall planning performance while reducing unnecessary rollout, and the authors report it transfers as a plug-in across different World Action Model architectures.
Why it matters
Autonomous-driving planners built on World Action Models spend compute imagining future scenes before acting, and treating every scene the same way wastes that budget on easy cases while potentially shortchanging risky ones. RISE's contribution is making the imagination length adaptive to estimated risk and expected benefit rather than fixed, which is the kind of incremental efficiency and safety improvement that tends to get folded into later planning stacks rather than shipped as a standalone product.
Who it affects
The direct audience is researchers and engineers building or evaluating World Action Models for autonomous driving, plus anyone working on safety-critical world-modeling who could reuse the CounterDrive dataset and its expert-verified counterfactual annotations. It has no direct bearing on drivers, riders, or products outside this research area.
How to use it
This is a research method and dataset rather than a released product, and the source gives no pricing, license, or availability details. The authors describe RISE as having plug-in generality, meaning it is designed to be layered onto different World Action Model architectures rather than tied to one system, and they describe CounterDrive as a reusable resource for other safety-critical world-modeling work.
How solid is it
The method is tested on two established autonomous-driving planning benchmarks, NAVSIM and nuScenes, and the CounterDrive samples are described as going through expert verification rather than being used raw. The material available here does not include the underlying accuracy or scoring numbers, the authors' names or institutional affiliations, the size of CounterDrive, or a publication venue or date, so none of those can be checked independently from this text.
Risks and caveats
The claims of best overall planning performance and plug-in generality across architectures come from the authors' own description of their experiments, and no baseline comparison numbers are present in the available text to weigh those claims against. The identity and criteria of the experts who verify CounterDrive samples are also not specified here.